Short Simulators, Long Experiments: A Controlled Audit of Horizon Mismatch in Simulation-Based Inference for Live-Cell Microscopy

Amortized simulation-based inference (SBI) may be trained on short synthetic recordings and then applied to consecutive windows of longer experiments. Unlike independent training clips, later windows inherit molecular composition, spatial correlations, and photophysical history. We quantify this horizon mismatch in a reaction--diffusion model of receptor dimerization with synthetic live-cell microscopy. For each of 500 parameter settings, we compare ten independent 2\,s simulations with ten consecutive 2\,s windows from one 20\,s simulation, using exact simulator truth. We examine the dimer-complex fraction, dimer unbinding rate, and monomer diffusion coefficient. By the final window, errors in the first two exceed prespecified tolerances, whereas diffusion remains within tolerance. For dimer composition, 90\% interval coverage falls from 0.95 to 0.39 without wider posterior intervals. Averaging windows reduces total error but not the history-induced error. We use its simulated magnitude as a sensitivity test, not a correction, for experimental MET recordings. The difference between resting and activated conditions remains much larger than this effect. However, absolute fractions remain model-dependent, and the workflow does not support interpreting a whole-recording unbinding estimate as a single constant rate. Short-window SBI should therefore be validated separately for each inferred quantity and aggregation rule.